International J ournal of Computer Trends and Technology- volume3Issue2- 2012

ISSN: 2231-2803 http://www.internationaljournalssrg.org Page 258

Performance analysis of Linear appearance based
algorithms for Face Recognition
Steven Lawrence Fernandes
Asst. Professor,
Dept of E & C,
Sree Devi Institute of Technology,
Mangalore, India

AbstractAnalysing the face recognition rate of various current
face recognition algorithms is absolutely critical in developing
new robust algorithms. I n his paper we propose performance
analysis of Principal Component Analysis (PCA), Linear
Discriminant Analysis (LDA) and Locality Preserving
Projections (LPP) for face recognition. This analysis was carried
out on various current PCA, LDA and LPP based face
recognition algorithms using standard public databases. Among
various PCA algorithms analyzed, Manual face localization used
on ORL and SHEFFIELD database consisting of 100
components gives the best face recognition rate of 100%, the
next best was 99.70% face recognition rate using PCA based
I mmune Networks (PCA-IN) on ORL database. Among various
LDA algorithms analyzed, I llumination Adaptive Linear
Discriminant Analysis (I ALDA) gives the best face recognition
rate of 98.9% on CMU PIE database, the next best was 98.125%
using Fuzzy Fisherface through genetic algorithm on ORL
database. Among various LPP algorithms analyzed, Subspace
Discriminant LPP (SDLLP) provides the best face recognition
rate of 98.38% on ORL database, the next best was 97.5% using
Contourlet-based Locality Preserving Projection (CLPP) on
ORL database.
Keywords-face recognition; Principal Component Analysis;
Linear Discriminant Analysis; Locality Preserving Projections;
PCA-I mmune Network; I llumination Adaptive LDA; Fisher
Discriminant; Subspace Discriminant LPP; Contourlet-based
Locality Preserving Projection.
I. INTRODUCTION
Facial recognition methods can be divided into appearance-
based or model-based algorithms. Appearance-based methods
represent a face in terms of several raw intensity images. An
image is considered as a high-dimensional vector. Statistical
techniques are usually used to derive a feature space fromthe
image distribution. The sample image is compared to the
training set.
Appearance methods can be classified as linear or non-
linear. Linear appearance-based methods perform a linear
dimension reduction. The face vectors are projected to the
basis vectors, the projection coefficients are used as the
feature representation of each face image, and approaches are
PCA, LDA, and LPP. Non-linear appearance methods are
more complicate. Linear subspace analysis is an
approximation of a nonlinear manifold. Kernel PCA (KPCA)
[23] is a method widely used.
Model-based approaches can be 2-Dimensional or 3-
Dimensional. These algorithms try to build a model of a
human face. These models are often morphable. A morphable
model allows classifying faces even when pose changes are
present, and approaches are Elastic Bunch Graph Matching
[24] or 3D Morphable Models [25].
Important information may be contained in the high-order
relationship among pixels, so Independent Component
Analysis [26-27] seems feasible to be a promising face feature
extraction method but ICA algorithms are iterative, time
consuming and have converge difficultly.
PCA is a transformation that chooses a new coordinate
system for data set that the greatest variance by any projection
of the data set comes to lieon the first axis (the first principal
component), the second axis corresponds to the maximum
remaining variations in the dimension orthogonal to the first
axis, and so on. The fundamental idea behind PCA is that if
there are series of multidimensional data vectors representing
objects which have similarities, it is possible to use
transformation matrix to create a reduces space that accurately
describes the original multidimensional vectors.
In LDA the original data is transformed into a lower
dimensional space (feature space) such that the ratio of the
between-class scatter matrix to within-class scatter matrix is
maximized i.e. the between-class scatter matrix is maximized
while the within-class scatter matrix is minimized and thus
maximumdiscrimination is achieved.
LPP proposed by He and Niyogi is an alternative to PCA.
LPP is a linear manifold learning approach. It can be viewed
as a linear approximation of Laplacian eigenmaps. The first
step of LPP is to generate an unsupervised neighborhood
graph on training data, and then finds an optimal locality
preserving projection matrix under certain criterion.
In this paper we report performance analysis of various
current PCA, LDA and LPP based algorithms for face
recognition. The evaluation parameter for the study is face
recognition rate on various standard public databases. The
remaining of the paper is organized as follows: Section II
provides a brief overview of PCA, Section III presents PCA
algorithms analysed, Section IV provides brief overview of
LDA, Section V presents LDA algorithms analysed. Section
VI provides brief overview of LPP, Section VII presents LPP
algorithms analysed. Section VIII presents performance
analysis of various PCA, LDA and LPP based algorithms
finally Section IX draws the conclusion.
International J ournal of Computer Trends and Technology- volume3Issue2- 2012

ISSN: 2231-2803 http://www.internationaljournalssrg.org Page 259

II. PRINCIPAL COMPONENT ANALYSIS (PCA)
Consider the training sample set of face image
} {
1 2
, , , ,
M
F x x x = .

where
( )
, , 1, ,
n
i i
x x R i M e = . corresponds to the
lexicographically ordered pixels of the ith face image, and
where there are M face images. PCA tries to mapp the
original n-dimensional image space into an mdimensional
feature space, where m<<n. The new feature vectors
m
i
y R e are defined by the following linear transform:
T
k k
y W x = 1, , k M = .

where
1 2
[ , , , ], ,
n
m i
W w w w w R = . e which is
orthogonal with each other is the eigenvector of total scatter
matrix
T
S corresponding to the
th
m largest eigenvalue. The
total scatter matrix is defined as
( )( )
1
M
T
T k k
k
S x x
=
=

Where is the mean value of all training samples.
III. PCA ALGORITHMS ANALYZED
A. PCA and Support Vector Machine (SVM)
PCA is used to extract the essential characteristics of face
images, SVM as classifier. One against one classification
strategy for multi-class pattern recognition is used based on
2D static face image [1].
B. Incremental Two-Dimensional Two-Directional
Principal Component Analysis (I(2D)2PCA)
Feature extraction method that combines advantages of
Two-Directional Principal Component Analysis (2D)
2
PCA
and Incremental PCA (IPCA). I(2D)
2
PCA consumes less
computational load than IPCA as well as smaller memory
waste than (2D)
2
PCA [2].
C. Infrared face recognition based on the Compressive
Sensing (CS) and PCA
The facial image is normalized and then the normalized
image does fast compressive sensing. PCA is used for non-
adaptive linear projections from CS which then classifies the
image using 3-nearest neighbor method [3].
D. Symmetrical Weighted Principal Component
Analysis (SWPCA)
SWPCA applies mirror transformto facial images, and
gets the odd and even symmetrical images based on the odd-
even decomposition theory. The Weighted Principal
Component Analysis (WPCA) is performed on the odd and
even symmetrical training sample sets respectively to extract
facial image features and nearest neighbor classifier is
employed for classification [4].
E. PCA based Immune Networks (PCA-IN)
PCA is utilized to obtain eigenvalues and eigenvectors of
the face images, and then the randomly selected single
training sample is input into the immune networks which are
optimized using genetic algorithms [5]. This experiment is
repeated for 30 times and the Average Recognition Rate
(ARR) is obtained.
F. Manual face localization
Localizes the face and eliminates the background
information fromthe image in a manner that the majority of
the cropped image consists of the facial pattern. Curvelet
transformis used to transformthe image into a new domain
and to calculate initial feature vectors. The feature vectors are
then dimensionally reduced using Two Dimensional Principal
Component Analysis (B2DPCA) and classified using
Extreme Learning Machine (ELM) [9].
G. Fractional Fourier transform(FRFT) and PCA
The face images are transformed into FRFT domain. PCA
is adopted to reduce the dimension of face images and
Mahalanobis distance is used for classifying [10].
H. Supervised learning framework for PCA-based face
recognition using Genetic Network Programming (GNP)
fuzzy data mining (GNP-FDM)
Genetic based Clustering Algorithm (GCA) is used to
reduce the number of classes. A Fuzzy Class Association
Rules (FCARs) based classifier is applied to mine the
inherent relationships between eigen-vectors [11].
I. PCA and minimumdistance classifier
Different facial images of a single human face are taken
together as a cluster. PCA is applied for feature extraction.
Minimumdistance classifier is used for the recognition that
avoids the exploit of threshold value which is changeable
under different distance classifiers [12].
IV. LINEAR DISCRIMANT ANALYSIS
Let us consider a set of N sample images {x
1
, x
2
,..., x
n
}
taking in an n-dimensional image space, and assume that each
image belongs to one of c classes {c
1
,c
2
,c
3
,,c
c
}.
Let Ni be the number of the samples in class
( )
1
1
1, 2, , ,
N
i i i
i
c i c x
N
=
= . . =


be the mean of the samples in class Ci. Then the between-
class scatter matrix Sb is defined as
( ) ( )
1
1
c
T
b i i i
i
S N
N
=
=

The within-class matrix S
w
is defined as
( ) ( )
1
1
k i
c
T
i i i i
i x c
S x x
N

= e
=


In LDA, the projection W
opt
is chosen to maximize the
ratio of the determinant of the between-class scatter matrix of
the projected samples to the determinant of the within-class
scatter matrix of projected samples
| |
1 2 3
arg
T
b
opt d
T
W S W
W mxw
W S W

= = .

where {w
i
|i=1,2, .,d} is the set of generalized eigen-
vectors of S
b
and S
w
corresponding to the m largest
generalized eigenvalues {
i
=1,2,,d}, i.e.,
International J ournal of Computer Trends and Technology- volume3Issue2- 2012

ISSN: 2231-2803 http://www.internationaljournalssrg.org Page 260

b i i i
S S

e = e
,
1,2, , . i d = ..

V. LDA ALGORITHMS ANALYZED
A. Regularized-LDA (R-LDA)
R-LDA is used for extracting low-dimensional discriminant
features fromhigh dimensional training images and then these
features are used by Probabilistic Reasoning Model (PRM) for
classification [13].
B. Multi-Feature Discriminant Analysis (MFDA)
Feature extraction method that combines advantages of
Two-Directional Principal Component Analysis (2D)
2
PCA
and Incremental PCA (IPCA). I(2D)
2
PCA consumes less
computational load than IPCA as well as smaller memory
waste than (2D)
2
PCA [2].
C. Rearranged Modular 2DLDA (Rm2DLDA)
Two-dimensional linear discriminant analysis has lower
time complexity but it implicitly avoids the small sample
problem encountered in classical LDA Rm2DLDA was
developed. It was based on the idea of dividing an image into
sub-images and then concatenating them to form a wide
image matrix [15].
D. Illumination Adaptive Linear Discriminant Analysis
(IALDA)
The images of many subjects under the different lighting
conditions are used to train illumination direction classifier
and varieties of LDA projection matrices. Then the
illumination direction of a test sample is estimated by
illumination direction classifier, the corresponding LDA
feature which is robust to the illumination variation between
images under the standard lighting conditions and the
estimated lighting conditions is extracted [16].
E. Fuzzy Fisherface (FLDA) through genetic algorithm
Searches for optimal parameters of membership function.
The optimal number of nearest neighbors to be considered
during the training is also found through the use of genetic
algorithms [17].
F. Semi-supervised face recognition algorithmbased
on LDA self-training)
Augments a manually labeled training set with new data
from an unlabeled auxiliary set to improve recognition
performance [18]. Without the cost of manual labeling such
auxiliary data is often easily acquired but is not normally
useful for learning.
G. Randomsampling LDA
To reduce the influence of unimportant or redundant
features on the variables generated by PCA, randomsampling
LDA was introduced. By incorporating Feature Selection for
face recognition (FS_RSLDA) was introduced, in this
algorithmunimportant or redundant features are removed at
first, this way the obtained weak classifier is made better
[19].
H. Revised Non-negative Matrix Factorization (NMF)
with LDA based color face recognition)
Block diagonal constraint is imposed on the base image
matrix and coefficient matrix on the basis of the constraints
of traditional NMF. And LDA is then implemented on
factorization coefficients to fuse class information [20].
I. Layered Linear Discriminant Analysis (L-LDA)
Decrease False Acceptance Rate (FAR) by reducing the
face dataset to very small size through L-LDA .It is intensive
to both small subspace (SSS) and large face variations due to
light or facial expressions by optimizing the separability
criteria. Hence it provides significant performance gain,
especially on similar face database and Small Subspace (SSS)
problems [21].
VI. LOCALITY PRESERVING PROJECTIONS
LPP is an unsupervised manifold dimensionality reduction
approach, which aims to find the optimal projection matrix W
by minimizing the following criterion function:
( )
2
1
,
.
i j ij
i j
J W y y S =

,
Substituting
T
i i
y W x =
into above objection function, it
yields by direct computation that
1
( ) 2 ( ),
T T
J W traceW XLX W =

Where L=D-S is called Laplacian matrix, D is a diagonal
matrix defined by
{
11 22
, , , }
NN
D diag D D D =

where .
ii ij
i
D S =

Matrix D provides a natural measure on the data points.
The bigger the value D
ii
(corresponding to y
i
) is, the more
important is y
i
. Thereby, the projection matrix W should
maximize the following constraint objective function
simultaneously:
2
1
( ) ( )
N
T T
i i
i
J W y Dy traceY DY
=
= =

( ).
T T
traceW XDX W =

Solving problems min J
1
(W) and max J
2
(W)
simultaneously is equivalent to minimizing the following
criterion function:
( )
( ) ,
( )
T T
T T
traceW XLX W
J W
traceW XDX W
=

The optimal locality preserving projection matrix
argmin ( )
d l
LPP
W R
W J W

e
=

can be obtained by solving the following generalized
eigen-value problem:
( ) .
T T
XLX W XDX W = A

where is a diagonal eigenvalue matrix. While in most
cases, the number of training data is smaller than the
dimension of feature vector, i.e. d << N. When this 3S
problemoccurs, the matrix
T
XDX is not full rank.
International J ournal of Computer Trends and Technology- volume3Issue2- 2012

ISSN: 2231-2803 http://www.internationaljournalssrg.org Page 261

VII. LPP ALGORITHMS ANALYZED
A. Bilateral Two-Dimensional LPP
It is based on 2D image matrices rather than column
vectors so the image matrix does not need to be transformed
into a long vector before feature extraction. The advantage
arising in this way is that 2D image matrices can be
effectively compressed from horizontal and vertical
directions and uses F-normclassification measure [30].
B. Regularized LPP (RLPP)
Supervised graph and regularization technique which not
only overcomes the singularity problemof LPP but also find
optimal LPP projection matrix in the entire input space [31].
C. Support Vector Machines (SVM) and LPP
It is combined semi-supervised face recognition, which
constructs data model on video sequence to unearth the
space-time information. It then uses semi-supervised LPP to
build a nearest neighbour graph which models the inherent
geometrical structure of the face space. SVM is applied to
find a separating hyperplane for the training data set and
predict the labels for the testing face data [32].
D. Subspace Discriminant LPP (SDLLP)
Based on discriminant graph and modified LPP criterion,
SDLLP method tackles the 3S issue of LPP and gives an
optimal algorithm by introducing corresponding eigen-
systems [33]. The projection feature vectors learned are linear
independent and make modified LPP criterion function reach
maximum.
E. Discrete sine transform(DST)
DST-feature based LPP algorithm is used for face
recognition and to obtain the 2D-DST facial feature. 1D-DST
is performed along two directions of a facial image matrix,
namely horizontal and vertical directions. A column DST-
feature vector is then formed by scanning its 2D-DST
coefficients along a zigzag route starting from the top left
corner and subsequently LPP can be performed in DST
domain directly [34].
F. Contourlet-based Locality Preserving Projection
(CLPP)
For dimensionality reduction and feature extraction CLPP
is used. The main advantage of CLPP over LPP lies in the
fact that the former realize dimensionality reduction
effectively than later. This can reduce the complexity greatly
and avoid the disadvantages of the pre-processing PCA step
[35].
G. Enhanced Supervised Locality Preserving
Projections (ESLPP)
An extension to LPP named ESLPP allows both locality
and class label information to be incorporated. This improves
the performance of classification. ESLPP uses similarity
based on robust path instead of Gaussian heat kernel
similarity. It can capture the underlying geometric
distribution of samples even when there are noise and outliers
[36].
H. Two-Dimensional Discriminant Locality Preserving
Projection (2DDLPP)
2DDLPP is a 2D-based feature extraction method, which
can directly process 2D image matrix without PCA pre-
processing. Hence it can be possible to retain the valid
information of human face images. 2DDLPP aims to preserve
the local manifold structure information and the discriminant
information [37].
I. Fractal Locality Preserving Projections (FLPP)
FLPP calculates the fractal codes of face images, and then use
LPP method to analysis the manifold structure, and do face
recognition [38].
J . Multilinear Principal Component Analysis (MPCA) and
LPP
It involves preprocessing in order to improve the face
image, then the feature extraction is achieved by merging
MPCA along with LPP. To calculate the feature projection
matrices, the test face sample is mapped onto a feature matrix
with the assistance of MPCA, LPP and there by the face
recognition is done by comparing the test feature matrix with
the enrolled face features in the database using L2 distance
[39].
K. Regularised Generalised Discriminant Locality
Preserving Projections (RGDLPP)
RGDLPP locality preserving within-class scatter is
replaced in DLPP approach by locality preserving total
scatter. To alleviate the problemof unreliable small and zero
eigenvalues caused by noise and the limited number of
training samples regularising the small and zero eigenvalues
of locality preserving within-class scatter is done, this enables
RGDLPP to be executed in the full sample space and
alleviates the over-fitting problem[40].
L. Tensor Locality Preserving Projections (TLLP)
TLLP is a natural extension of LPP to the multilinear case,
which can take data directly in the form of tensors of
arbitrary order as input hence there exist significant reduction
in both space complexity and time complexity [41].
VIII. PERFORMANCE ANALYSIS
A. Performance analysis of various PCA based algorithms
Illumination invariant face recognition based on DCT and
PCA on YALE Database B gives accuracy of 94.2% [28].
TABLE I. PERFORMANCE COMPARISON BETWEEN
PCA+NN, SVM AND PCA+SVM ON ORL DATABASE
Class
Number
Training
samples
Test
samples
Method
Recognition
rate (%)

200 C

60

140
PCA+NN 90
SVM 85.71
PCA+SVM 94.29

40 C

120

280
PCA+NN 80.36
SVM 78.93
PCA+SVM 81.10
As Table I shows, face recognition rate of PCA+SVM
method, under small samples circumstance, is better than
PCA+NN and SVM [1].
International J ournal of Computer Trends and Technology- volume3Issue2- 2012

ISSN: 2231-2803 http://www.internationaljournalssrg.org Page 262

TABLE II. PERFORMANCE COMPARISON BETWEEN PCA,
2DPCA, (2D)2PCA, IPCA, I(2D)PCA AND I(2D)2PCA ON YALE
DATABASE
Method Recognition rate (%)
PCA 80.80
2DPCA 82.05
(2D)
2
PCA 82.13
IPCA 78.47
I(2D)PCA 81.19
I(2D)
2
PCA 81.39
TABLE III. PERFORMANCE COMPARISON BETWEEN PCA,
2DPCA, (2D)2PCA, IPCA, I(2D)PCA AND I(2D)2PCA ON ORL
DATABASE
Method Recognition rate(%)
PCA 85.14
2DPCA 86.29
(2D)
2
PCA 86.64
IPCA 84.75
I(2D)PCA 86.16
I(2D)
2
PCA 86.28
Table II and Table III shows, face recognition rate of
I(2D)
2
PCA is better when compared to PCA, 2DPCA,
(2D)
2
PCA, IPCA, I(2D)PCA on YALE and ORL databases
[2].
TABLE IV. PERFORMANCE COMPARISON BETWEEN
EIGENFACE, EIGEN-GEFES AND EIGEN-GEFEW ON FRGC
DATABASE
Methods Recognition rate (%)
Eigenface 87.14
Eigen-GEFeS 86.67
Eigen-GEFeW 91.42
Table IV shows, Eigen-GEFeW is the best performing
instance when compared with Eigenface and Eigen-GEFeS
on Face Recognition Grand Challenge (FRGC) dataset [29].
Infrared face recognition based on the compressive sensing
and PCA is invariant to variations in facial expressions and
viewpoint, and is computationally efficient [3].
TABLE V. PERFORMANCE COMPARISON BETWEEN PCA,
SPCA, WPCA AND SWPCA ON ORL DATABASE
Method
Training
samples/class
Recognition rate
(%)
PCA 6 92.50
SPCA 6 94.37
WPCA 6 94.37
SWPCA 6 96.00
TABLE VI. PERFORMANCE COMPARISON BETWEEN PCA,
SPCA, WPCA AND SWPCA ON YALE DATABASE
Method
Training
samples/class
Recognition rate
(%)
PCA 4 85.71
SPCA 4 88.57
WPCA 4 89.52
SWPCA 4 93.33
Table V and Table VI shows, the correct recognition
accuracy with SWPCA improved almost by 10% compared
with PCA. The reason that the SWPCA method performs
better than other conventional algorithms is that SWPCA not
only utilizes the natural symmetrical property of human face
to enlarge the number of training samples, but also employs
the weighted PCA space to improve the robustness against
variance of illumination and expression [4].
TABLE VII. PERFORMANCE COMPARISON BETWEEN
2DPCA, DDCT AND 2PCA, MODULAR WEIGHTED (2D)2PCA AND
PCA-IN ON ORL DATABASE
Methods Recognition rate (%)
2DPCA [6] 76.70
DDCT and 2PCA [7] 76.22
Modular Weighted
(2D)
2
PCA [8]
72.22
PCA-IN 99.70
Table VII shows, best performance (99.70%) of PCA-IN
classifiers was compared with the results reported in [6-8].
PCA-IN method outperformed all other methods [5]. Face
recognition rate of Manual face localization on ORL and
SHEFFIELD database consisting of 100 components is 100%
[9].
In FRFT face images are transformed into FRFT domain,
it uses several angles characters for classifying. Experiments
on FERET database shows that FRFT provides new insights
into the role that pre-processing methods play in dealing with
images [10].
GNP-FDM successfully prevents the accuracy loss caused by
a large number of classes in the Multiple Training Images per
Person Complicated Illumination Database (MTIP-CID). GCA
reduces the overlaps in the PCA domain [11].
PCA and minimumdistance classifier gives a recognition
rate of 96.7% on ORL database [12].
B. Performance analysis of various LDA based algorithms
TABLE VIII. PERFORMANCE COMPARISON BETWEEN R-LDA
AND R-LDA USING PRM ON YALE DATABASE
Number of
features
Methods
Recognition rate
(%)
32 R-LDA 95
32
R-LDA
Using
PRM
97.5
TABLE IX. PERFORMANCE COMPARISON BETWEEN R-LDA
AND R-LDA USING PRM ON UMIST DATABASE
Number of
features
Methods
Recognition rate
(%)
12 R-LDA 88.50
12
R-LDA Using
PRM
98.48
International J ournal of Computer Trends and Technology- volume3Issue2- 2012

ISSN: 2231-2803 http://www.internationaljournalssrg.org Page 263

Table VIII and Table IX shows, R-LDA Using PRM gives
better recognition when compared to R-LDA on YALE and
UMIST databases. Further it is observed that by taking more
number of features (32), the recognition rate is maximum
(97.5%) for ORL database and by considering 12 number of
features in case of UMIST database the recognition rate is
98.48% [13].
Compared to LDA, MFDA significantly boosts the
recognition performance. The accuracy for LDA is 60%
compared to the 83.9% accuracy of MFDA [14].
TABLE X. PERFORMANCE COMPARISON BETWEEN
2DLDA, RM2DLDA (2X2) AND RM2DLDA (4X4) ON ORL DATABASE
Methods Recognition rate (%)
2DLDA 95.65
Rm2DLDA(2 x 2) 96.65
Rm2DLDA(4 x 4) 97.1
TABLE XI. PERFORMANCE COMPARISON BETWEEN
2DLDA, RM2DLDA (2X2) AND RM2DLDA (4X4) ON YALEB
DATABASE
Methods Recognition rate (%)
2DLDA 88.68
Rm2DLDA(2 x 2) 90.75
Rm2DLDA(4 x 4) 91.55
TABLE XII. PERFORMANCE COMPARISON BETWEEN
2DLDA, RM2DLDA (2X2) AND RM2DLDA (4X4) ON PIE DATABASE
Methods Recognition rate (%)
2DLDA 90.58
Rm2DLDA(2 x 2) 93.0
Rm2DLDA(4 x 4) 95.04
Table X, Table XI and Table XIII shows, Rm2DLDA
gives better recognition when compared to 2DLDA on ORL,
YALE and PIE databases [15].
TABLE XIII. PERFORMANCE COMPARISON BETWEEN LDA
AND IALDA ON B1 DATABASE
Method
Training
samples/class
Test
samples/class
Recognition rate
(%)
LDA 2 43 59.38
IALDA 1 44 85.52
TABLE XIV. PERFORMANCE COMPARISON BETWEEN LDA
AND IALDA ON CMU PIE DATABASE
Method
Training
samples/class
Test
samples/class
Recognition rate
(%)
LDA 2 19 74.25
IALDA 1 20 98.9
Table XIII and Table XIV show, IALDA gives better
recognition when compared with LDA on B1 and CMU
databases [16]. The recognition rate is increased from94.12%
using Fuzzy Fisherface (FLDA) to 98.125% using Fuzzy
Fisherface through genetic algorithmon ORL database [17].
Experiments on ORL database, AR database and CMU
PIE database show that Semi-supervised face recognition
algorithm based on LDA is robust to variations in
illumination, pose and expression and that it outperforms
related approaches in both transductive and semi-supervised
configurations [18].
RSLDA is an effective random sampling LDA method, the
1-NN classifier in the feature subspace obtained by RSLDA
has better classification performance as compared to that
induced by BaseLDA on AR, ORL, YALE, YALEB face
datasets [19]. For ORL, the classification accuracy has an
increase of 15.1% around.
Experimental results on CVL and CMU PIE databases
prove the algorithm improves recognition rate effectively
[20]. L-LDA is insensitive to large dataset and also small
sample size and it provided 93% accuracy and reduced False
Acceptance Rate (FAR) to 0.42 on BANCA face database
[21].
C. Performance analysis of various LPP based algorithms
B2DLPP can better avoid the interference of light, noise
and other factors via extracting discriminate information.
Also dimension reduction by two directions needs less cost of
space and time than that by one side, so eigenvector space is
far smaller than the original space and it can avoid a SSS [30].
TABLE XV. PERFORMANCE COMPARISON BETWEEN DLPP
AND RLPP ON ORL DATABASE
Method
Training
samples
Recognition rate
(%)
DLPP 5 89.65
RLPP 5 96.80
TABLE XVI. PERFORMANCE COMPARISON BETWEEN DLPP
AND RLPP ON FERET DATABSE
Method
Training
samples
Recognition rate
(%)
DLPP 5 78.17
RLPP 5 92.00
Table XV shows, the correct recognition accuracy with
DLPP and RLPP is 89.65% and 96.80% with training number
5 respectively for ORL database. Table XVI shows, the
correct recognition accuracy with DLPP and RLPP is 78.17%
and 92.00 % with training number 5 respectively for FERET
database [31].
Support Vector Machines and Locality Preserving
Projections can not only discover the spatio-temporal
connection between video faces, but also take advantage of
the superiority of semi-supervised learning and manifolds
learning [32].
TABLE XVII. PERFORMANCE COMPARISON BETWEEN DLPP,
PCA+LPP AND SDLPP ON ORL DATABSE
Method
Training
samples
Recognition rate
(%)
DLPP 8 92.75
PCA+LPP 8 95.63
SDLPP 8 98.38
International J ournal of Computer Trends and Technology- volume3Issue2- 2012

ISSN: 2231-2803 http://www.internationaljournalssrg.org Page 264

TABLE XVIII. PERFORMANCE COMPARISON BETWEEN DLPP,
PCA+LPP AND SDLPP ON FERET DATABASE
Method
Training
samples
Recognition rate
(%)
DLPP 8 81.75
PCA+LPP 8 79.92
SDLPP 8 91.58
Table XVII and Table XVIII show, SDLPP provides better
recognition when compared to DLPP and PCA+LPP on ORL
and FERET databases [33].
TABLE XIX. PERFORMANCE COMPARISON BETWEEN LPP,
OLPP, WPCA AND CLPP ON ORL DATABSE
Method
Recognition rate
(%)
LPP 92.0
OLPP 95.5
CLPP 97.5
Table XIX shows, CLPP provide better recognition when
compared to LPP and OLPP on ORL database which contains
images from 40 individuals, each providing 10 different
images [34].
TABLE XX. PERFORMANCE COMPARISON BETWEEN LPP,
OLPP, WPCA AND CLPP ON YALE DATABSE
Method
Recognition rate
(%)
LPP 90.5
OLPP 92.0
CLPP 94.1
Table XX shows, CLPP provide better recognition when
compared to LPP and OLPP on YALE database which
contains 165 images of 15 individuals [34].
TABLE XXI. PERFORMANCE COMPARISON BETWEEN LPP,
OLPP, WPCA AND CLPP ON PIE DATABSE
Method
Recognition rate
(%)
LPP 90.5
OLPP 90.5
CLPP 95.8
Table XXI shows, CLPP provide better recognition when
compared to LPP and OLPP on PIE database of 41,368
images of 68 people [34].
TABLE XXII. PERFORMANCE COMPARISON BETWEEN
LaplacianfaceAND DST-LPP ON ORL DATABSE
Method
Training
samples
Recognition rate
(%)
Laplacianface 5 96.30
DST-LPP 5 96.60
TABLE XXIII. PERFORMANCE COMPARISON BETWEEN
LaplacianfaceAND DST-LPP ON FERET DATABSE
Method
Training
samples
Recognition rate
(%)
Laplacianface 5 89.67
DST-LPP 5 92.33
Table XXII and Table XXIII show, DST-LPP provides
better recognition when compared to Laplacianface on ORL
and FERET databases respectively with 5 training number
[35].
TABLE XXIV. PERFORMANCE COMPARISON BETWEEN LPP,
OLPP AND ESLPP ON ORL DATABSE
Method
Training
samples
Recognition rate
(%)
LPP 5 90.50
OLPP 5 91.00
ESLPP 5 92.45
TABLE XXV. PERFORMANCE COMPARISON BETWEEN LPP,
OLPP AND ESLPP ON YALE DATABSE
Method
Training
samples
Recognition rate
(%)
LPP 5 88
OLPP 5 91.50
ESLPP 5 93.50
Table XXIV and Table XXV show, ESLPP provides better
recognition when compared to LPP and OLPP on ORL and
YALE databases respectively with 5 training number [36].
TABLE XXVI. PERFORMANCE COMPARISON BETWEEN LPP
AND 2DDLDPP ON ORL DATABASE
Method
Training
samples
Recognition rate
(%)
LPP 5 90.58
2DDLDPP 5 96.45
Table XXVI show, 2DDLDPP provides better recognition
when compared to LPP on ORL database with 5 training
number [37].
TABLE XXVII. PERFORMANCE COMPARISON BETWEEN LPP
AND 2DDLDPP on YALE database
Method
Training
samples
Recognition rate
(%)
LPP 8 93.33
2DDLDPP 8 96.56
Table XXVII show, 2DDLDPP provides better recognition
when compared to LPP on YALE database with 5 training
number [37].
TABLE XXVIII. PERFORMANCE COMPARISON BETWEEN LPP
AND FLPP ON ORL DATABASE
Method Recognition rate (%)
LPP 94.3
FLPP 95.4
TABLE XXIX. PERFORMANCE COMPARISON BETWEEN LPP
AND FLPP ON YALE DATABASE
Method Recognition rate (%)
LPP 86.2
FLPP 86.5
International J ournal of Computer Trends and Technology- volume3Issue2- 2012

ISSN: 2231-2803 http://www.internationaljournalssrg.org Page 265

Table XXVIII and Table XIX show, FLPP provides better
recognition when compared to LPP and on ORL and YALE
databases respectively with 5 training number [38].
TABLE XXX. PERFORMANCE COMPARISON BETWEEN
MPCA+LDA AND MPCA+LPP ON FERET DATABASE
Methods Recognition rate (%)
MPCA+LDA 93.75
MPCA+LPP 96.5
TABLE XXXI. PERFORMANCE COMPARISON BETWEEN
MPCA+LDA AND MPCA+LPP ON AT&T DATABASE
Method Recognition rate (%)
MPCA+LDA 95
MPCA+LPP 96.5
Table XXX and Table XXXI show, MPCA+LPP provides
better recognition when compared to MPCA+LDA and on
FERET and AT&T databases respectively [39].
RGDLPP statistically significantly outperforms LPP,
DLPP algorithms on PIE database [40].
TABLE XXXII. PERFORMANCE COMPARISON BETWEEN LPP,
TLPP ON ORL DATABASE
Method Recognition rate (%)
LPP 96.6
TLPP 97.1
Table XXXII show, LPP provides better recognition when
compared to TLPP on ORL database [41].
D. Performance comparison between PCA and LDA based
algorithms
TABLE XXXIII. PERFORMANCE COMPARISON BETWEEN
EIGENFACES AND FISHERFACES ON YALE DATABASE
Number of
features
Methods
Recognition rate
(%)
32 Eigenfaces 90.5
32 Fisherfaces 93.5
Table XXXIII shows, LDA gives better recognition when
compared to PCA while 32 features are considered on YALE
Database [13].
TABLE XXXIV. PERFORMANCE COMPARISON BETWEEN
EIGENFACES AND FISHERFACES ON UMIST DATABASE
Number of
features
Methods
Recognition rate
(%)
12 Eigenfaces 90.62
12 Fisherfaces 94.45
Table XXXIV shows, LDA gives better recognition when
compared to PCA while 12 features are considered on
UMIST Database [13].
TABLE XXXV. PERFORMANCE COMPARISON BETWEEN
2DPCA AND RLDA ON ORL DATABASE
Methods Recognition rate (%)
2DPCA 77.86
RLDA 73.89
Table XXXV shows, 2DPCA gives better recognition
when compared to RLDA on ORL Database [15]
TABLE XXXVI. PERFORMANCE COMPARISON BETWEEN
2DPCA AND RLDA ON YALEB DATABASE
Methods
Recognition rate
(%)
2DPCA 77.86
RLDA 73.89
Table XXXVI shows, 2DPCA gives better recognition
when compared to RLDA on YALEB Database [15].
TABLE XXXVII. PERFORMANCE COMPARISON BETWEEN
2DPCA AND RLDA ON PIE DATABASE
Methods
Recognition rate
(%)
2DPCA 87.74
RLDA 92.22
Table XXXVII shows, RLDA gives better recognition
when compared to 2DPCA on PIE Database [15].
TABLE XXXVIII. PERFORMANCE COMPARISON BETWEEN PCA
AND LDA ON B1 DATABASE
Method
Training
samples/class
Test
samples/class
Recognition rate
(%)
PCA 1 44 57.2
LDA 2 43 59.38
TABLE XXXIX. PERFORMANCE COMPARISON BETWEEN PCA
AND LDA ON CMU PIE DATABASE
Method
Training
samples/class
Test
samples/class
Recognition rate
(%)
PCA 1 20 64.56
LDA 2 19 74.25
Table XXXVIII and Table XXXIX show, LDA gives
better recognition when compared to PCA on B1 and CMU
PIE Databases respectively [16].
TABLE XXXX. PERFORMANCE COMPARISON BETWEEN PCA
AND LDA ON ATT, CROPPED YALE, FACES94, FACES95, FACES96,
J AFE DATABASES
Database Name LDA PCA
ATT 94.40 91.30
CROPPED YALE 93.80 90.30
FACES95 90.80 87.00
FACES96 97.20 94.00
FromTable XXXX, it is evident that the best algorithmto
recognize image without disturbance is PCA, because in the
same recognition rate, PCA takes shorter time than LDA. But
to recognize image with disturbances, LDA is better to use
because it has better recognition rate [22]. In termof time
taken, PCA tends to be much better than LDA, especially to
recognize images with background disturbance [22].
International J ournal of Computer Trends and Technology- volume3Issue2- 2012

ISSN: 2231-2803 http://www.internationaljournalssrg.org Page 266

IX. CONCLUSION
In this paper, we have analysed various current PCA, LDA
and LPP based algorithms for face recognition. This analysis
is vital in developing new robust algorithms for face
recognition. Among various PCA algorithms analysed, the
best result was found when Manual face localization was
used on ORL and SHEFFIELD database consisting of 100
components. The face recognition rate in this case was 100%.
The next best was 99.70% face recognition rate using PCA-
IN on ORL database. Among various LDA algorithms
analysed, it was found that IALDA gives the best face
recognition rate of 98.9 % when 20 test samples and 1
training sample were considered on CMU PIE Database. The
next best was 98.125 % using Fuzzy Fisherface through
genetic algorithm on ORL database. Among various LPP
algorithms analyzed, Subspace Discriminant LPP (SDLLP)
provides the best face recognition rate of 98.38% on ORL
database, the next best was 97.5% using Contourlet-based
Locality Preserving Projection (CLPP) on ORL database.
REFERENCES
[1] Chengliang Wang, Libin Lan, Yuwei Zhang, and Minjie Gu, "Face
Recognition Based on Principle Component Analysis and Support
Vector Machine," 2011 3rd International Workshop on Intelligent
Systems and Applications, pp. 1 - 4, May 2011.
[2] Yonghwa Choi, Tokumoto, T., Minho Lee, and Ozawa, S.,
"Incremental two-dimensional two-directional principal component
analysis (I(2D)2PCA) for face recognition," 2011 IEEE International
Conference on Acoustics, Speech and Signal Processing, pp. 1493 -
1496, May 2011.
[3] Zhang Lin, Zhang Wenrui, Sun Li, and Fang Zhijun, "Infrared face
recognition based on compressive sensing and PCA," 2011 IEEE
International Conference on Computer Science and Automation
Engineering, vol. 2, pp. 51 - 54, J une 2011.
[4] Guoxia Sun, Liangliang Zhang, and Huiqiang Sun, "Face Recognition
Based on Symmetrical Weighted PCA," 2011 International
Conference on Computer Science and Service System, pp. 2249 -
2252, August 2011.
[5] Guan-Chun Luh, "Face recognition using PCA based immune
networks with single training sample per person," 2011 International
Conference on Machine Learning and Cybernetics, vol. 4, pp. 1773 -
1779, J uly 2011.
[6] J ian Yang, David Zhang, Alejandro Frangi, and Jing-yu Yang, Two-
dimensional PCA: a new approach to appearance-based face
representation and recognition, IEEE Transactions on Pattern
Analysis and Machine Intelligence, Vol. 26, No. 1, pp. 131-137, J an.
2004.
[7] J un-ying Gan, J ian-hu Gao, and Chun-zhi Li, Face recognition based
on diagonal discrete cosine transformand two dimensional principal
component analysis, Computer Engineering and Applications, Vol.
43, No. 31, pp. 210-213, 2007.
[8] Dashun Que, Bi Chen and Jin Hu, A Novel SingleTraining Sample
Face Recognition AlgorithmBased on Modular Weighted (2D)2PCA,
Proceedings of ICSP Conference, Beijing, pp. 1552-1555, Oct. 2008
[9] A. Adeel Mohammed, Rashid Minhas, Q. M. J onathan Wu, and M. A.
Sid-Ahmed, "A face portion based recognition system using
multidimensional PCA," 2011 IEEE 54th International Midwest
Symposiumon Circuits and Systems, pp. 1 - 4, August 2011.
[10] Zhoufeng Liu, and Xiangyang Lu, "Face recognition based on
fractional Fourier transform and PCA," 2011 Cross Strait Quad-
Regional Radio Scienceand Wireless Technology Conference, vol. 2,
pp. 1406 - 1409, J uly 2011.
[11] Deng Zhang, Shingo Mabu, Feng Wen, and Kotaro Hirasawa, "A
Supervised Learning Framework for PCA-based Face Recognition
using GNP Fuzzy Data Mining," 2011 IEEE International Conference
on Systems, Man, and Cybernetics, pp. 516 - 520, October 2011.
[12] Bag Soumen, and Sanyal Goutam, "An Efficient Face Recognition
Approach using PCA and Minimum Distance Classifier," 2011
International Conference on Image Information Processing, pp. 1 - 6,
November 2011.
[13] Lingraj Dora, N. P. Rath, "Face Recognition by Regularized-LDA
Using PRM," 2010 International Conference on Advances in Recent
Technologies in Communication and Computing, pp. 140 - 145,
October 2010.
[14] Zhifeng Li, Unsang Park, and J ain, A.K, "A Discriminative Model for
Age Invariant Face Recognition," IEEE Transactions on Information
Forensics and Security, vol. 6, no. 3, pp. 1028 - 1037, September 2011.
[15] Huxidan J umahong, Wanquan Liu, and Chong Lu, "A new rearrange
modular two-dimensional LDA for face recognition," 2011
International Conference on Machine Learning and Cybernetics, vol. 1,
pp. 361 - 366, J uly 2011.
[16] Zhonghua Liu, J ingbo Zhou, and Zhong J in, "Face recognition based
on illumination adaptive LDA," 2010 20th International Conference
on Pattern Recognition, pp. 894 - 897, August 2010.
[17] Amar Khoukhi, Syed Faraz Ahmed, "Fuzzy LDA for Face
Recognition with GA Based Optimization," 2010 Annual Meeting of
the North American Fuzzy Information Processing Society, pp. 1 - 6,
J uly 2010.
[18] Xuran Zhao, Nicholas Evans, and J ean-Luc Dugelay, "Semi-
supervised face recognition with LDA self-training," 2011 18th IEEE
International Conference on Image Processing, pp. 3041 - 3044,
September 2011.
[19] Ming Yang, J ian-Wu Wan, and Gen-Lin J i, "Randomsampling LDA
incorporating feature selection for face recognition," 2010
International Conference on Wavelet Analysis and Pattern
Recognition, pp. 180 - 185, J uly 2010.
[20] Xiaoming Bai, and Chengzhang Wang, "Revised NMF with LDA
based color face recognition," 2010 2
nd
International Conference on
Networking and Digital Society, vol. 1, pp. 156 - 159, May 2010.
[21] Muhammad Imran Razzak, Muhammad Khurram Khan, Khaled
Alghathbar, and Rubiyah Yousaf, "Face Recognition using Layered
Linear Discriminant Analysis and Small Subspace," 2010 IEEE 10th
International Conference on Computer and Information Technology,
pp. 1407 - 1412, J uly 2010.
[22] Erwin Hidayat, Fajrian Nur A., Azah Kamilah Muda, Choo Yun Huoy,
and Sabrina Ahmad, "A Comparative Study of Feature Extraction
Using PCA and LDA for Face Recognition," 2011 7th International
Conference on Information Assurance and Security, pp. 354 - 359,
December 2011.
[23] Chandar K.P., Chandra M.M., Kumar M.R., Latha, B.S., "Multi scale
feature extraction and enhancement using SVD towards secure Face
Recognition system," 2011 International Conference on Signal
Processing, Communication, Computing and Networking
Technologies, pp. 64 - 69, J uly 2011.
[24] Pervaiz, A.Z., "Real Time Face Recognition SystemBased on EBGM
Framework," 2010 12th International Conference on Computer
Modelling and Simulation, pp.262 - 266, March 2010.
[25] Weyrauch, B., Heisele, B., Huang, J ., and Blanz, V., "Component-
Based Face Recognition with 3D Morphable Models," 2004
CVPRW'04 Conference on Computer Vision and Pattern Recognition
Workshop, pp.85, J une 2004.
[26] Zhao Lihong, Wang Ye, and Teng Hongfeng, "Face Recognition
based on Independent Component Analysis," 2011 Chinese Control
and Decision Conference, pp. 426 - 429, May 2011.
[27] J iajin Lei, Chao Lu, and Zhenkuan Pan, "Enhancement of
Components in ICA for Face Recognition," 2011 9th International
Conference on Software Engineering Research, Management and
Applications, pp. 33 - 38, August 2011.
[28] Shermina J , "Illumination invariant face recognition using Discrete
Cosine Transform and Principal Component Analysis," 2011
International Conference on Emerging Trends in Electrical and
Computer Technology, pp. 826 - 830, March 2011.
[29] Abegaz, T., Dozier, G., Bryant, K., Adams, J ., Popplewell, K., Shelton,
J ., Ricanek, K., and Woodard, D.L., Hybrid GAs for Eigen-based
facial recognition, 2011 IEEE Workshop on Computational
International J ournal of Computer Trends and Technology- volume3Issue2- 2012

ISSN: 2231-2803 http://www.internationaljournalssrg.org Page 267

Intelligence in Biometrics and Identity Management, pp. 127 - 130,
April 2011.
[30] J iadong Song, Xiaojuan Li, Jinhua Zhong, Pengfei Xu, and Mingquan
Zhou, A Novel Face Recognition Method: Bilateral Two
Dimensional Locality Preserving Projections (B2DLPP), IEEE-
Pattern Recognition (CCPR), 2010 Chinese Conference on Pattern
Recognition, Oct. 2010, pp. 15
[31] Wen-Sheng Chen, Wei Wang, and J ian-Wei Yang, A Novel
Regularized Locality Preserving Projections for Face Recognition,
IEEE-Information and Computing (ICIC), 2011 Fourth International
Conference on Intelligent Computing, April 2011, pp. 110113
[32] Ke Lu, Zhengming Ding, J idong Zhao, and Yue Wu, A novel Semi-
supervised Face Recognition for Video, IEEE-Intelligent Control
and Information Processing(ICICIP), 2010 International Conference
on Intelligent Control and Information Processing, Aug 2010, pp.
313316
[33] Wei He, Wen-Sheng Chen, and Bin Fang, A Novel Subspace
Discriminant Locality Preserving Projections for Face Recognition,
IEEE-Computational Intelligence and Security(CIS), 2011 Seventh
International Conference on Computational Intelligence and Security,
Dec 2011, pp. 10571061
[34] Yanqi Tan, Yaying Zhao, and Xiaohu Ma, Contourlet-Based Feature
Extraction with LPP for Face Recognition, IEEE-International
Conference on Multimediaand Signal Processing(CMSP), May 2011,
pp. 122125
[35] Wei Wang and Wen-Sheng Chen, DST Feature based Locality
Preserving Projections for Face Recognition, IEEE-International
Conference on Computational Intelligence and Security(CIS), Dec
2010, pp. 288292
[36] Xian-Fa Cai, Gui-Hua Wen, J ia Wei, and Jie Li, Enhanced
supervised locality preserving projections for face recognition, IEEE-
International Conference on Machine Learning and
Cybernetics(ICMLC), J uly 2011, pp. 17621766
[37] Xiajiong Shen, Qing Cong, and Sheng Wang, Face Recognition
Based on Two-Dimensional Discriminant Locality Preserving
Projection, IEEE-3rd International Conference on Advanced
Computer Theory and Engineering(ICACTE), Aug 2010, pp. V4-403-
V4-407
[38] Yang Liu and Jin-guang Sun, Face Recognition Method Based on
FLPP, IEEE-Third International Symposium on Electronic
Commerce and Security(ISECS), J uly 2010, pp. 298301
[39] Shermina.J , Face recognition system using Multilinear Principal
Component Analysis and Locality Preserving Projection, IEEE-GCC
Conference and Exhibition (GCC), Feb 2011, pp. 283286
[40] Lu, G.-F., Lin, Z., and J in, Z., Face recognition using regularised
generalised discriminant locality preserving projections, IEEE-
Computer Vision, IET, March 2011, pp. 107116
[41] Dazhao Zheng, Xiufeng Du, and Limin Cui, Tensor Locality
Preserving Projections for Face Recognition, IEEE-International
Conference on Systems Man and Cybernetics(SMC), Oct 2010, pp.
23472350